[davidmckayv]

Every Bot's working rules, reworked against what dogfooding showed: replies that ran to 600 words for a yes-or-no question, "are you able to...?" answered by doing unasked work, "done" said before anything was checked, a Python script approved on a computer with no Python, and spreadsheets left plain.

What changes for every Bot

• One push to act, bounded by what was asked. The push to act was said three times (<how_to_work>, the ask test's "keep going" and the Personal Assistant's role) with nothing about scope. It is now said once, in <how_to_work>: do the preparation a task needs without asking, take a decided step in the same reply, and do not start work nobody asked for, so a question gets its answer, not the job it is about. The ask test's else-branch now reads "carry on with what was asked".
• A search that has not come back. A Bot that started a tool search waits for it before deciding, or saying, that something cannot be done.
• Failures by kind. A temporary failure gets one retry; a declined step stays declined; a change whose outcome is unknown is looked for in the app before it is redone; a failure that keeps happening becomes a short report.
• Claims need a tool result. "Done, sent, saved or correct" only when a tool result in this conversation showed it (Anthropic's own rule for agents handling records). After a change in an app, read it back once. Report scoped to what was checked; "all" only if every item was checked.
• Worded per model. Anthropic's Opus 5 guide says that model reads its own work back unprompted and over-verifies when told to, so readsBack leaves the read-back sentence out for Claude Opus 5 and later; the claims rule stays for every model. GPT-6.x and Claude Sonnet/Haiku keep a bounded "once", which is OpenAI's own fix for GPT-6 testing too broadly. Gemini 3.x has no such guidance and keeps it.
• <replies>, new: size the reply to the question and lead with the answer. Yes/no gets yes or no and at most one line more. A question about whether the Bot can do something gets a short answer from what it holds and an offer, not the work, not a read of the person's data, not a capability list. A request worded as a question ("can you send me today's agenda?") is a request. A finished job is one or two sentences with the result or a link. No recaps, no repeated summaries, no message per step. It says it is about words: a component the answer calls for is still drawn, so generative UI is unchanged.
• <documents>, new: spreadsheets get labelled input cells, live formulas, units in headers, a bold frozen header row, real number, percent, currency and date formats, one tab for a small task, kept formatting, and a read-back for #REF!, #NAME? and #VALUE! before delivery. Documents get a title, headings and no filler.
• What the computer has. computerGuidance takes an optional installed list from the deployment that builds the computer's image, and tells the Bot to check before writing a script and never to ask the person to approve a script for a program the computer does not have.
• The "do it now" nudge lets an offer stand. Its words now hold the Bot to work it was asked for, and leave an offer for the person.
• The Personal Assistant's role drops the repeated push to act and the duplicate ask test; it keeps what is particular to an assistant and adds "when they ask whether you can, answer and offer" and "keep replies short".

Tests

• shared/bot-prompt.test.ts: the push to act appears once across every Bot's working rules and "keep going" nowhere; scope, search wait and typed retries; the claims rule; the read-back present for GPT-5.6, GPT-6.1, Sonnet 5.5, Haiku 5.5, Opus 4.8, Gemini 3.8 and unknown models, absent for Opus 5 and later; reply shapes including the request-worded-as-a-question case and the line that keeps components; spreadsheet defaults; the computer's software line only when given; section order; no shouting.
• agent-claude-sdk/tests/fixtures/server-instructions.json regenerated, and test_openbot_instructions.py passes against it.
• The server tests that pin the section order and the nudge's words are updated to the new ones.

Evidence from real-model runs, before and after, is below.